hagrid

Why we included this project

Hand gesture recognition training data tends to be small or synthetic; HaGRIDv2 is a notable exception. It packs over a million FullHD RGB images across 33 gesture classes plus a separate no-gesture class, all with bounding-box annotations, so you can train both classifiers and detectors. The collection spans nearly 66,000 individuals filmed at distances from half a meter to four meters, with varied indoor lighting, which gives models a realistic spread of conditions. Splitting by person rather than random frames reduces test leakage, and the repo includes download tooling plus a companion algorithm that learns dynamic gestures from these static images.